South Africa Does Not Need an AI Skills Pipeline. It Needs an AI Capability System

By Dr Riaan Steenberg

Ten thousand young people complete an AI course. They receive certificates. The completion rate is reported. The programme is described as a contribution to employability.

Six months later, we still do not know whether they can use AI to perform valuable work, whether employers changed any jobs to use their skills, or whether a meaningful number entered employment.

This is the difference between a skills pipeline and a capability system.

South Africa is right to invest in artificial-intelligence education. The draft national AI policy calls for AI learning from school through higher education, stronger industry partnerships, educator development, interdisciplinary capability and wider access. The Department of Higher Education and Training has entered partnerships that promise large-scale digital and AI training opportunities.

These are useful inputs. They are not yet an employment system.

A country does not become capable because many people complete courses. It becomes capable when learning is converted into reliable performance, performance is connected to real demand, and people are given a route through which they can become useful.

The Pipeline Metaphor Is Too Simple

The pipeline metaphor assumes that people enter education, acquire skills and flow into jobs waiting at the other end.

South Africa's labour market does not behave like that.

In the second quarter of 2026, the official unemployment rate reached 33.6%. Youth unemployment for people aged 15 to 34 rose to 47.4%. The country does not have a simple shortage of trained people feeding an abundant supply of appropriate work. It has mismatches between education, location, networks, experience, employer demand, economic growth and the structure of opportunity.

Adding an AI certificate to that system may improve an individual's prospects. It does not repair the system.

The pipeline also encourages institutions to count what is easiest to count: enrolments, completions, certificates and training hours. Those measures show activity. They do not establish capability, utilisation or employment.

The relevant sequence is harder:

learning, practice, observed performance, trusted evidence, work opportunity, contribution, feedback and progression.

That is not a pipeline. It is a system with loops.

AI Literacy Is Necessary but Insufficient

Most workers do not need to become machine-learning engineers. The OECD notes that fewer than one percent of workers are likely to need advanced AI skills. Far more people need digital competence, the ability to work with data, and the judgement to use AI appropriately. Problem-solving, creativity, managerial ability and other human capabilities remain important.

This should change how South Africa frames the skills problem.

Basic AI literacy matters. People should understand what generative systems can do, where they fail, how to protect sensitive information, how bias enters a result and why verification matters. They should be able to formulate a useful request, inspect an output and recognise when the tool is unsuitable.

But literacy is not capability.

A person may know how to use an AI assistant and still be unable to perform an accounting reconciliation, design a learning intervention, diagnose a production fault, interpret customer evidence or make a defensible management recommendation. The tool can amplify domain capability. It cannot supply all of the context, standards and judgement that competent work requires.

The useful unit is therefore not "AI skill" in isolation. It is AI applied to a real domain under real conditions.

Capability Requires a Context

Skills become valuable inside a context of work.

Consider a student who completes a course in prompt design. What can an employer reasonably infer? The person may know how to elicit a structured answer from a language model. That says little about whether they can detect a fabricated source, protect personal information, understand a client problem, select an appropriate method or accept responsibility for a recommendation.

Now consider a logistics trainee who uses AI to investigate late deliveries. They must understand route data, operational constraints, customer promises, exceptions, cost and escalation. They use the tool to find patterns, but they test those patterns against actual records. A supervisor reviews the analysis. The trainee explains which recommendation they rejected and why. The result is observed over time.

That is capability.

It combines knowledge, tools, judgement, standards, evidence and consequences. It develops through performance and correction, not content exposure alone.

This is why industry partnership must mean more than companies donating course licences. Industry must help define valuable work, supply authentic problems, create supervised practice and recognise credible evidence of performance.

The Missing Middle Is Practice

South Africa often separates education from experience and then asks young people to solve the gap themselves.

Employers advertise entry-level positions that require experience. Graduates search for internships to become eligible for jobs for which they have already studied. Training programmes teach concepts without access to the systems, data and operating conditions in which those concepts matter.

AI may widen this gap because it raises the apparent standard of entry-level output. A graduate with intelligent tools can produce a professional-looking report. Employers may then expect polished work from the first day while remaining reluctant to provide the supervision through which judgement develops.

The result is a dangerous illusion: output looks more capable while the development pathway becomes thinner.

A capability system must deliberately create the missing middle. It needs workplace simulations, apprenticeships, supervised projects, virtual practice environments, community problem laboratories and employer-backed challenges. Learners should encounter incomplete information, weak data, conflicting priorities and the need to explain a decision.

The aim is not merely to demonstrate that AI can complete the task. The aim is to develop a person who can govern the use of AI in completing it.

Evidence Must Travel With the Person

Certificates are useful because they compress information. An employer cannot personally inspect every learning experience. The certificate acts as a signal that a standard was met.

In a fast-changing field, that signal needs support.

A capability record should show what the person did, under which conditions, with which tools, against which standard and with what result. It should include assessed artefacts, observed demonstrations, explanations, revisions and supervisor judgements. Where work was AI-assisted, the record should make clear what the person contributed and how the output was verified.

This is not a call for an enormous digital surveillance system. It is a call for better evidence.

The evidence should be portable enough to help a learner cross institutional boundaries. A TVET student, university graduate, employed worker or self-taught practitioner should be able to demonstrate capability without relying solely on the reputation of the institution that trained them.

The stronger the evidence, the less employers need to use background, accent, networks or institutional prestige as crude proxies for potential.

Employers Must Create Demand, Not Only Specify It

The skills conversation often treats employer demand as a fact that education must discover. Employers list desired competencies, institutions adjust curricula and young people are told to become employable.

But employers also design the demand.

They decide which tasks are automated, which remain human, how work is grouped into roles, whether entry-level positions exist, how much supervision is available and whether productivity gains are used to develop people or reduce headcount. They decide whether AI creates a wider doorway into work or a higher barrier at the entrance.

An AI capability system therefore requires employers to redesign jobs, not merely report skills shortages.

They should identify tasks that can be performed by developing workers with appropriate tools and review. They should create progression from assisted work to independent judgement. They should measure whether training improves performance and mobility, not merely whether staff attended it.

If industry wants work-ready graduates, industry must help provide the work through which readiness develops.

Public-Private Partnerships Need Outcome Logic

Technology companies can bring useful platforms, curricula, instructors and scale. Public institutions can bring reach, legitimacy and integration into national systems. Partnerships between them can expand access quickly.

But every partnership should carry an explicit outcome logic.

Who is the target learner? What opportunity is the programme meant to unlock? Which capability will be demonstrated? Who recognises the evidence? Where will practice occur? Which employers are committed to interviews, projects, placements or redesigned roles? What happens to learners who complete the programme but cannot enter work? Which outcomes will be measured after six, twelve and twenty-four months?

Without those questions, the public announcement becomes the outcome. Scholarships offered are treated as people trained. People trained are treated as people capable. Certificates awarded are treated as people employed.

Each step requires different evidence.

This is not an argument against scale. It is an argument that scale should apply to the whole conversion system, not only the easiest input.

What an AI Capability System Contains

South Africa already has many of the institutions required: schools, universities, TVET and community colleges, SETAs, professional bodies, employers, technology companies, research organisations and public employment services. The problem is not the absence of parts. It is the weakness of the connections between them.

A functioning capability system would contain at least seven elements:

1. Foundational access. Connectivity, devices, language support and basic digital competence.
2. Domain-linked learning. AI taught through accounting, teaching, healthcare, manufacturing, logistics, agriculture and public administration-not as a detached novelty.
3. Supervised practice. Real or simulated work with review, feedback and increasing complexity.
4. Competence evidence. Portable proof of what a person can do, including how AI was used and governed.
5. Employer participation. Work design, projects, placements, apprenticeships and explicit recognition of the evidence.
6. Transition support. Career guidance, networks, matching, transport or data support, and help crossing from training into work.
7. Outcome measurement. Employment, earnings, retention, mobility, productivity and enterprise creation measured over time.

None of these can be inferred from course completion alone.

A Better National Question

The draft AI policy correctly recognises that South Africa needs talent development, infrastructure, governance, inclusion and industry collaboration. The implementation challenge is to avoid translating those ambitions into a catalogue of training initiatives.

The country should not ask only how many people can be trained in AI. It should ask which economically and socially valuable capabilities can be expanded, where demand for them will exist, and how people will produce trusted evidence that they can perform.

The difference matters because unemployment is not a content deficit. A person can learn more and remain excluded. A training programme can succeed administratively and fail economically. A certificate can be genuine and still carry little labour-market value.

South Africa needs AI skills. But skills become productive only when they are joined to practice, evidence, opportunity and a route to progression.

Build that system, and training becomes one of its most important components.

Build only the pipeline, and we may become very good at delivering certificates into an economy that still cannot use the people holding them.

Sources

• Statistics South Africa, Quarterly Labour Force Survey Q2 2026
• South Africa, Draft National Artificial Intelligence Policy (2026)
• Department of Higher Education and Training and Google South Africa AI and digital-skills memorandum (2026)
• OECD, AI and Skills (2026)
• World Economic Forum, Artificial Intelligence and the Future of Entry-Level Work (2026)